Global crop yields

This release is built from the merged agricultural database. HarvestStat Asia takes precedence where a usable regional layer exists, followed by GMFD / Hultgren regional records, then FAOSTAT national records. We select one entire reporting layer per country and crop. We never splice tiers across years or overlay national values on regional values. The source and resolution selectors can override this default. Some regional records are older than national records; use “Latest observed harvest” to see their actual end years.

Yield-only sources do not require annual production or area. Fixed reference crop areas are metadata, not annual observations or aggregation weights. We do not manufacture country yields by averaging regional yields. Eligible source observations retain zeros and upstream warnings; upstream errors and ambiguous duplicate years are excluded. All original rows, year labels, quality flags, and corrections remain downloadable in each regional panel.

Known rice forms use a fixed paddy-equivalent conversion: milled weights divided by 0.67, brown weights by 0.80, paddy unchanged. Areas are unchanged. These factors affect absolute yields, not percentage responses from a log-yield model. Rice of unspecified weight basis remains searchable and usable for relative ENSO responses, but is excluded from comparable absolute-yield maps. Spring and winter wheat remain separate seasons; an explicit season selection reveals each.

Local default seasons are chosen per reporting unit, preferring an annual series, then the season with the most eligible observations. We never pool seasons or source products. A fixed period retains the same requested years everywhere. “Available record” uses each source series’ actual span. Gray means unavailable or filtered, not zero yield. Twenty-six source units lack a safe boundary match; their records remain accessible through the location list.

Niño 3.4 and the relative index come from the merged database’s ERSSTv5 series, with a 1991–2020 anomaly baseline. The relative index subtracts tropical-mean SST and is unscaled; it is not official RONI. MEI.v2 uses NOAA PSL’s JRA3Q product in standardized index units, with overlapping two-month periods. We average the MEI periods whose end month lies in the selected window, so its first period includes the preceding month. The observed-data right axis always shows native index units. New index products can be added through the versioned registry without changing the agricultural observations.

ENSO models fit log yield with a linear time term and linear, quadratic, or separate warm/cold index slopes. All registered indices and models use the same positive-yield years with complete exposure; MEI coverage limits these comparisons to 1979 onward. A fit requires 20 years and five exposures on each side of zero. The default period is 1981–2024. A standard-deviation scenario uses the index variability in the fitted sample; a native scenario uses the selected index’s units. A +3°C event profile is available for SST indices only and is not converted into MEI. These are associations and illustrative scenarios, not yield forecasts.

Yield trend is a separate regression of observed yield on time, reported in t/ha per decade. Both trend and ENSO intervals use 400 shared three-year pairs-bootstrap samples. ENSO intervals are normal intervals in log response transformed to percent, pointwise without a field multiplicity correction. Validation withholds five-year blocks under two layouts and excludes adjacent training years. Positive holdout skill requires improvement over a trend-only model in both layouts. Fit details report sensitivity to a quadratic time trend and restoration of statistical-outlier-only records.

Coverage catalog · Import audit · Source lineage and corrections · Index registry and values · Release checksums · Methods and reproduction · Earlier nine-country atlas

Country–crop files and fits are compressed JSON. Each regional panel provides CSV downloads of original observations, inclusion flags, fitted values, and selected scenarios. Boundaries are simplified for display. Source attribution: HarvestStat Asia, Hultgren et al. (2025), FAOSTAT, crop calendars, and boundary providers. Inputs retain their upstream terms.